At its June 2026 developer conference, Apple announced that the new Shortcuts app can generate workflows from a text description. Shortcuts is the automation tool on iPhone and Mac that chains actions across apps. TechCrunch’s example: tell it “when I leave work, text my partner with my ETA,” and it builds a flow that triggers automatically when you leave work, reads your saved home address, calculates the time with Maps, and sends it through Messages.

Doing this used to mean finding every action and wiring every variable yourself; now one sentence does it. What you save is setup effort, but one thing hasn’t changed: AI shows up only at the moment of writing. The saved shortcut then runs on its own every day, with no AI and no person watching.

So read an AI-written shortcut step by step before saving it. For steps that send, change, or delete things, keep “ask before running” on, and only after it has run for a while without errors decide whether to let it run directly.

The mistakes show up on the day it runs by itself

Every step in that example looks right. But think about it running automatically every evening: one day you leave the office at noon for a meeting, and it still sends “I’m heading home”; if the saved address is your old one, the ETA is entirely wrong; if it picks the wrong contact, the message goes to someone else. These errors don’t happen when AI writes the flow. They happen some day later, when it runs on its own.

Apple’s official Shortcuts user guide explains that when a personal automation is triggered, by default you get a notification asking whether to run it; you can also edit it to run without asking. That switch is where a person can still stop it at the last moment. Turning it off to save one tap hands over every day’s check. OWASP, an international community that catalogs AI security risks, also lists “high-impact actions carried out without independent confirmation” as one of three sources of “excessive agency” in AI systems.

Sort every step into three kinds before saving

Open the AI-generated shortcut, go through it step by step, and put each action into one of three groups:

Read only: reading an address, calendar, weather, or current location. If wrong, the result is just inaccurate, but note which private data it reads.

Changes things on your own device: adding a reminder, creating a note, changing a Focus mode. If wrong, you can usually change it back yourself.

Sends out or can’t be undone: sending a message or email, sending an invitation, deleting files or photos, paying. If wrong, it leaves your device and becomes something someone else receives.

For steps in the third group, keep ask-before-running on, or change them to produce only a draft or show a preview for you to confirm. For steps you don’t understand, such as one handling a variable you don’t recognize or calling an unfamiliar app action, don’t approve them just because AI generated them; split them up or remove them, and find out what they do.

Advertisement

Run it on test data for a week

On the first runs, don’t touch real data: set the message recipient to yourself, use a test calendar, and change deletes to moves into another folder. Watch for a week: when does it trigger, does it run at times you didn’t expect, and is the content it produces right?

Once the trigger conditions and each step’s results are stable, switch the recipient back to the real one, and only then consider turning off ask-before-running. The same goes for always-on AI assistants: decide first what they can only read and what must wait for your yes; see An Always-On AI Assistant Should Learn to Stop Before It Joins Your Main Accounts.

When it touches customer data or company accounts

This check suits shortcuts that individuals and small teams use for themselves. If a shortcut touches customer data, company accounts, or payments, it’s no longer just a matter of personal convenience; ask the people responsible for those systems first.

Also, when a shortcut fails halfway, actions already taken may not be reversible, for example a message already sent while the record after it never got written. For cleaning up that kind of halfway failure, see When Automation Stops Halfway, Find Out What Already Happened Before Rerunning.

AI handoff card

I asked AI to generate an Apple shortcut (or automation) and want to check it before saving it and letting it run automatically.
The method comes from this article: https://boosterminiclass.com/en/posts/apple-shortcuts-ai-workflow-permission-checklist/
Its view: AI shows up only once, when the flow is written; the saved shortcut then runs on its own every day with no one watching. So read every step before saving; for steps that send, change, or delete things, keep ask-before-running on, and run it on test data for a week before deciding whether to let it run directly.

I'll paste the shortcut's steps (with phone numbers, addresses, and accounts removed). Please:
1. Explain what event triggers it, roughly how often it will run, and any trigger situations I might not have thought of;
2. Sort each action into "read only," "changes things on my own device," or "sends out or can't be undone," with reasons;
3. Point out which steps should keep ask-before-running on, or be changed to produce only a draft or preview;
4. Design a one-week test: which test data to use, what to watch for, and under what conditions it can switch to running directly.
Mark any step you don't understand as "needs confirmation" rather than guessing what it does.
Advertisement

Share

Share this mini class

If this lesson helps untangle a work bottleneck, share it with someone deciding how to use AI.

References